qwen3.8 27b + vllm + github copilot
Replaced original with yours, it is working much better, the original was breaking copilot all the time.
I told Qwen itself to compare both jinjas, your's much superior according to it 😀
Thanks.
It doesn't work for me unfortunately. When it tries to send back a tool call it fails with the XML such as:
<function=list_dir> <parameter=path> c:\src\ai-toolkit\docs\code-optimization <function=file_search> <parameter=query> **/krea* <parameter=maxResults> 50
And copilot then stops.
I tried overriding the output to JSON, but it still comes back as XML
Main difference is I am using LM Studio instead of VLLM
If you are using an lmstudio curated model the template comes bundled into the gguf. Use that instead.
Don't use LMStudio for serving models: behind the scenes it is llama.cpp but outdated, with some options missing and other options enforced even if they are not ideal/compatible.
You'll be much better off using llama-server directly.
That comes down to how LM Studio serves tool calls compared to vLLM.
GitHub Copilot expects the server to catch the tool call and send it back as a structured tool_calls JSON block in the API response.
When you run vLLM with --tool-call-parser qwen3_xml, it intercepts Qwen's <tool_call> XML and turns it into the JSON format Copilot needs. LM Studio doesn't have an XML tool parser on its local server, so it dumps the raw <tool_call> text straight into the chat content. Copilot doesn't recognize raw XML in the message body as a tool call, so it stops.
If you want to use Qwen with Copilot locally, running llama-server (with --tool-call-parser qwen3_xml) or vLLM is the way to go.